arsen
|
9433070864
|
feat(adr-119): MLP classifier (22→32→6) replaces LogReg fallback
Single-hidden-layer perceptron (~3k params, ReLU + softmax) trained via
manual backprop (no external ML crate). SGD + momentum 0.9 + weight
decay 1e-4 + cosine LR decay, 30 epochs over 151,329 frames.
AdaptiveModel carries both LogReg and MLP weights side-by-side;
classify() prefers MLP via is_trained() check, falls back to LogReg
when loading legacy 15-feature models.
Result on same 6-node 7-class dataset:
LogReg (ADR-118): 49.58%
MLP (this): 53.53% (+3.95 pts)
Per-class gains concentrated on motion classes — exactly where
non-linear feature combinations matter:
absent +1 (40% → 41%)
present_still tied (99% → 99%, class-imbalance ceiling)
transition +7 (29% → 36%)
active +8 (22% → 30%)
waving +4 (34% → 38%)
present_moving +9 (24% → 33%)
Cumulative session improvement vs 2-node 15-feature baseline:
40.4% → 53.53% (+13.1 pts).
Loss flatlines at 1.15 around epoch 10 — frame-level information
ceiling for the 22-feature representation. Next big lever is
temporal context (windowed LSTM/TCN), documented in Out-of-scope.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
|
2026-05-18 00:48:19 +07:00 |